The DAO soft-fork try was troublesome. Not solely did it end up that we underestimated the unintended effects on the consensus protocol (i.e. DoS vulnerability), however we additionally managed to introduce a knowledge race into the rushed implementation that was a ticking time bomb. It was not ideally suited, and though averted on the final occasion, the quick approaching hard-fork deadline regarded eerily bleak to say the least. We wanted a brand new technique…
The stepping stone in the direction of this was an concept borrowed from Google (courtesy of Nick Johnson): writing up an in depth postmortem of the occasion, aiming to evaluate the foundation causes of the problem, focusing solely on the technical points and applicable measures to forestall recurrence.
Technical options scale and persist; blaming folks doesn’t. ~ Nick
From the postmortem, one fascinating discovery from the angle of this weblog put up was made. The soft-fork code inside [go-ethereum](https://github.com/ethereum/go-ethereum) appeared strong from all views: a) it was totally coated by unit checks with a 3:1 test-to-code ratio; b) it was totally reviewed by six basis builders; and c) it was even manually stay examined on a non-public community… But nonetheless, a deadly knowledge race remained, which may have probably brought on extreme community disruption.
It transpired that the flaw may solely ever happen in a community consisting of a number of nodes, a number of miners and a number of blocks being minted concurrently. Even when all of these eventualities held true, there was solely a slight probability for the bug to floor. Unit checks can not catch it, code reviewers could or could not catch it, and handbook testing catching it might be unlikely. Our conclusion was that the event groups wanted extra instruments to carry out reproducible checks that might cowl the intricate interaction of a number of nodes in a concurrent networked state of affairs. With out such a instrument, manually checking the assorted edge instances is unwieldy; and with out doing these checks repeatedly as a part of the event workflow, uncommon errors would turn into unimaginable to find in time.
And thus, hive was born…
What’s hive?
Ethereum grew massive to the purpose the place testing implementations turned an enormous burden. Unit checks are positive for checking varied implementation quirks, however validating {that a} shopper conforms to some baseline high quality, or validating that purchasers can play properly collectively in a multi shopper setting, is all however easy.
Hive is supposed to function an simply expandable check harness the place anybody can add checks (be these easy validations or community simulations) in any programming language that they’re comfy with, and hive ought to concurrently be capable of run these checks towards all potential purchasers. As such, the harness is supposed to do black field testing the place no shopper particular inside particulars/state might be examined and/or inspected, reasonably emphasis could be placed on adherence to official specs or behaviors beneath totally different circumstances.
Most significantly, hive was designed from the bottom as much as run as a part of any purchasers’ CI workflow!
How does hive work?
Hive’s physique and soul is [docker](https://www.docker.com/). Each shopper implementation is a docker picture; each validation suite is a docker picture; and each community simulation is a docker picture. Hive itself is an all encompassing docker picture. It is a very highly effective abstraction…
Since Ethereum clients are docker pictures in hive, builders of the purchasers can assemble the very best setting for his or her purchasers to run in (dependency, tooling and configuration sensible). Hive will spin up as many situations as wanted, all of them working in their very own Linux methods.
Equally, as test suites validating Ethereum purchasers are docker pictures, the author of the checks can use any programing setting he’s most aware of. Hive will guarantee a shopper is working when it begins the tester, which might then validate if the actual shopper conforms to some desired habits.
Lastly, network simulations are but once more outlined by docker pictures, however in comparison with easy checks, simulators not solely execute code towards a working shopper, however can really begin and terminate purchasers at will. These purchasers run in the identical digital community and might freely (or as dictated by the simulator container) join to one another, forming an on-demand non-public Ethereum community.
How did hive assist the fork?
Hive is neither a substitute for unit testing nor for thorough reviewing. All present employed practices are important to get a clear implementation of any characteristic. Hive can present validation past what’s possible from a mean developer’s perspective: working intensive checks that may require complicated execution environments; and checking networking nook instances that may take hours to arrange.
Within the case of the DAO hard-fork, past all of the consensus and unit checks, we would have liked to make sure most significantly that nodes partition cleanly into two subsets on the networking degree: one supporting and one opposing the fork. This was important because it’s unimaginable to foretell what antagonistic results working two competing chains in a single community might need, particularly from the minority’s perspective.
As such we have carried out three particular community simulations in hive:
-
The first to examine that miners working the total Ethash DAGs generate appropriate block extra-data fields for each pro-forkers and no-forkers, even when making an attempt to naively spoof.
-
The second to confirm {that a} community consisting of combined pro-fork and no-fork nodes/miners accurately splits into two when the fork block arrives, additionally sustaining the cut up afterwards.
-
The third to examine that given an already forked community, newly becoming a member of nodes can sync, quick sync and light-weight sync to the chain of their alternative.
The fascinating query although is: did hive really catch any errors, or did is simply act as an additional affirmation that the whole lot’s all proper? And the reply is, each. Hive caught three fork-unrelated bugs in Geth, however additionally closely aided Geth’s hard-fork growth by repeatedly offering suggestions on how modifications affected community habits.
There was some criticism of the go-ethereum workforce for taking their time on the hard-fork implementation. Hopefully folks will now see what we have been as much as, whereas concurrently implementing the fork itself. All in all, I consider hive turned out to play fairly an necessary position within the cleanness of this transition.
What’s hive’s future?
The Ethereum GitHub group options [4 test tools already](https://github.com/ethereum?utf8=%E2percent9Cpercent93&question=check), with no less than one EVM benchmark instrument cooking in some exterior repository. They aren’t being utilised to their full extent. They’ve a ton of dependencies, generate a ton of junk and are very difficult to make use of.
With hive, we’re aiming to mixture all the assorted scattered checks beneath one common shopper validator that has minimal dependencies, might be prolonged by anybody, and might run as a part of the each day CI workflow of shopper builders.
We welcome anybody to contribute to the mission, be that including new purchasers to validate, validators to check with, or simulators to search out fascinating networking points. Within the meantime, we’ll attempt to additional polish hive itself, including assist for working benchmarks in addition to mixed-client simulations.
With a bit or work, perhaps we’ll even have assist for working hive within the cloud, permitting it to run community simulations at a way more fascinating scale.